Atlas-Based Multiorgan Segmentation for Dynamic Abdominal PET

Region of interest (ROI) delineation is required to extract tissue time-activity curves (TACs) in dynamic PET studies, to analyze functional changes or estimate physiological parameters. In this paper, we present an automatic framework for atlas-based multiorgan segmentation in abdominal dynamic PET images with three different methods (4D-pair, 4D-PCA, and 3-D), incorporating probabilistic atlas information into the segmentation as a spatial prior using maximum a posteriori (MAP) estimation. Due to different tracer kinetics in each organ, PET images from different time periods post injection (p.i.) have great intensity differences. Thus, when dynamic images are available, to use this temporal information, two strategies can be employed. First, tissue activities from two frames with highly different activity distributions were selected, namely, an early 8-10 min p.i. and a late 55-60 min p.i. frame, and modeled as a bivariate Gaussian distribution. Theoretically, this method can be applied if more than one frame of data is available. Second, principal component analysis (PCA) was applied to the full series of dynamic images to extract two images corresponding to the first two components. When dynamic image series are not available, the segmentation framework can be scaled down to 3-D, by building a univariate Gaussian distribution based on one 3-D image. The final segmentation results for all three methods were determined by optimizing the MAP-based energy function with two hyperparameters (λl, η) by multilabel graph cuts. We performed hyperparameter optimization and evaluated the proposed segmentation methods of 4D-pair and 4D-PCA by leave-one-out cross-validation using 30 sets of 4-D abdominal 18F-FP-(+)-DTBZ PET images. To evaluate segmentation results, the pancreas and spleen TACs were extracted, and the percentage error between the area under curve (AUC) of the TACs extracted by manual and automated segmentations was determined. The 4D-pair method with the hyperparameter combination of (λ_l = 0.1, η = 1) yielded the best performance. TAC AUC %error results with PCA-based methods showed slightly higher %error than 4D-pair. The 3-D method showed much larger %error than the other two methods. The 4D-pair results agreed well with the manual segmentation, with mean pancreas and spleen TAC AUC %errors of 0.3±3.3% and -0.4±8.1%, respectively. In addition, the distribution volume (V
_T
) values of pancreas and spleen were determined by kinetic modeling using TACs from either manual or automated segmentations. There were no significant differences between manual- and auto-V
_T
values (p values of 0.14 and 0.74 for pancreas and spleen, respectively). Thus, the proposed automated segmentation method can provide robust and reliable ROIs of the pancreas and spleen for kinetic modeling.